I put language models
behind deterministic guardrails.

You upload an Excel template; Voxa reads its schema, transcribes what you narrate with Whisper, extracts the fields with gpt-4o-mini, and appends the row. Running on AWS EC2 behind Caddy, with DNS, backups and CI/CD provisioned in Terraform.

The public demo spends real money on every call, so I contained it with hourly and daily per-IP limits over a persisted budget ledger — chosen instead of a sign-up wall so the demo stays frictionless. Releases carry no long-lived credentials and no inbound SSH: GitHub OIDC into AWS, multi-arch images to GHCR, rollout through SSM Run Command.

The 660 tests run with no network and no credentials — including one that asserts the offline property itself, written after a leaked environment variable let the suite pass silently for weeks.

A statistician narrates a play; the system resolves the game. The model extracts only what was actually said — an unmentioned field stays absent, never a courtesy zero — while a deterministic rules engine computes down and distance, possession, turnovers, yard-line math, scoring and penalties. Narration and derived state persist in separate columns so the two never blur.

Post-game corrections are safe by construction: because each play's state derives from the one before it, editing a play re-runs the engine across the rest of the game and rewrites every downstream state.

I built an evaluation harness over a golden narration set that scores per-field accuracy and contract violations, counting courtesy fills as their own failure mode. Latest run: 100 of 103 expected fields across 16 plays. On top sit role-scoped JWT auth and a 15-table catalog on versioned SQL migrations.

Building the second product exposed how much of the first was domain-agnostic. I extracted the transcription and extraction pipeline into an installable package — Whisper, the LLM call with retry and error translation, tolerant JSON parsing with sparse mode, the data contracts and the exception taxonomy — consumed through a single prompt-builder interface.

The seam is the point: product schema, vocabulary and prompts stay out of the engine, so the football domain never leaks into the generic path. Anything a second client would also need gets promoted upward, deliberately, and logged.

Architected an end-to-end pipeline consolidating three Stripe accounts and Airtable into a single PostgreSQL warehouse — Airbyte → dbt → REST API — serving the production dashboard. Built Go/Gin APIs for subscriptions, one-time purchases and investor–GP matching, backed by materialized views for sub-second reads from the React front end.

Designed and deployed the AWS footprint with Terraform (RDS, ECS Fargate for dbt runs, SSM Parameter Store) across dev and prod, and integrated Keycloak SSO with custom email-gating middleware. The dbt models unify customers, subscriptions and purchases under a 24-hour staleness contract agreed with the business — a number the business owns, not one I invented.

Designed an immutable attendance model in which corrections stack on top of the original punch instead of overwriting it, backed by an append-only audit trail enforced through Postgres permissions rather than application convention — so the log holds even when a handler is wrong.

Every attendance judgment is derived at read time: punches paired against the effective schedule (employee override, else area), shifts crossing midnight, forgotten shifts auto-closed. No rollup tables, so a late correction resolves without a backfill. Closed a check-then-insert race on punch initiation with a transactional advisory lock, so a resent one-time link reuses the pending request instead of spending a second biometric verification.

I also introduced the frontend test harness the project lacked and wrote its architecture decision log.

A team contribution, not a solo product: a team of three, code-reviewed by the technical lead. The ADR log runs to 41 entries — 27 are decisions I made, the rest I documented from existing code.

An ETL pipeline turning 1,620+ athletic evaluations scattered across Excel files into a validated, normalized PostgreSQL database covering 229+ athletes. Integrated the Model Context Protocol so coaching staff query that database in natural language through Claude, instead of waiting on someone to write the SQL.

Currently designing a custom MCP server with curated tools, read-only roles and audit logging — because natural-language access to a production database is only safe if the surface is deliberately small.

In progress: the MCP integration is in use; the custom server with its own tool surface is still being built.